Design of a rain rate retrieval algorithm using artificial neural

Comments

Transcription

Design of a rain rate retrieval algorithm using artificial neural
Revista Ingenierı́a UC, Vol. 23, No. 2, agosto 2016 153 - 161
Design of a rain rate retrieval algorithm using artificial neural network
and the advanced technology microwave sounder
Erith Muñoz∗,a , Francesco Di Paolab , Mario Lanfric
a
The Food and Agriculture Organization, Quito, Ecuador
b
National Research Council, Italy
c
Comisión Nacional de Actividades Espaciales, Instituto Mario Gulich, Córdoba, Argentina
Abstract.Passive microwave remote sensing techniques have shown to be a useful tool in order to retrieve atmospheric parameters
on basis of radiative transfer fundamentals. However, due to the difficulties related with both numerical and computing
complexity to model the physical phenomenology involved in atmospheric dynamics, some different techniques, such as
Artificial Neural Networks (ANN), have been tested to find the mathematical relationship between atmospheric state variables
and measurements carried out at a given time by satellite instruments. In this paper, it is presented a heavy rain-rate
retrieval algorithm (ANN183) developed by using both an artificial neural network trained through the L-BFGS algorithm
and measurements related with the strong water vapor absorption line at 183 GHz provided by the Advanced Technology
Microwave Sensor (ATMS) aboard of the National Polar-orbiting Partnership (Suomi-NPP) platform of the National Oceanic
and Atmospheric Administration (NOAA-USA). Preliminaries results showed a strong potential to reproduce the spatial
distribution of rainfall when is compared with estimates provided by gauges measurements.
Keywords: atmospheric remote sensing; rain rate retrieval algorithms; artificial neural networks; ATMS
Diseño de un algoritmo para estimación de tasa de precipitación usando
redes neuronales artificiales y la sonda de microondas de tecnologı́a
avanzada
Resumen.Las técnicas de teledetección usando microondas pasivas, han mostrado ser una herramienta útil para estimar parámetros
atmosféricos a partir de fundamentos de transferencia radiativa. Sin embargo, debido a las dificultades computacionales
asociadas con complejidades de orden numérico para modelar los fenómenos fı́sicos que tienen lugar en la dimámica
atmosférica, se han probado diversas técnicas, tales como redes neuronales artificiales, para definir relaciones matemáticas
funcionales entre el estado de las variables atmosféricas y las mediciones realizadas por intrumentos a bordo de plataformas
satelitales. En este trabajo, se presenta un algoritmo (ANN183) para estimar tasa de precipitación desarrollado a partir del
entrenamiento de redes neuronales mediante el método L-BFGS, y usando mediciones relacionadas con la fuerte lı́nea de
absorción por parte del vapor de agua a 183 GHz proporcionadas por la Sonda de Microondas de Tecnologı́a Avanzada (ATMS,
por sus siglas en inglés) a bordo de la plataforma de la alianza nacional de satélites polares Suomi-NPP de la Administración
Nacional del Océano y la Atmósfera (NOAA). Los resultados preliminares muestran un potencial relevante para reproducir la
distribución espacial de precipitaciones, lo cual fue evidenciado mediante comparaciones realizadas con datos proporcionados
por estaciones pluviométricas en superficie.
Palabras clave: teledetección atmosférica; estimación de tasa de precipitación; redes neuronales artificiales;
ATMS
Revista Ingenierı́a UC, ISSN: 1316–6832, Facultad de Ingenierı́a, Universidad de Carabobo.
E. Muñoz et al / Revista Ingenierı́a UC , Vol. 23, No. 2, agosto 2016, 153-161
154
Recibido: mayo 2016
Aceptado: julio 2016
1. Introducción
The brightness temperature (BT) at a given
atmospheric region might be linked to the electromagnetic density power measured by a radiometer through the radiative transfer equation
[1, 2]. However, instead of that, commonly the
microwave satellite rain-rate retrieval algorithms
are developed by defining a mathematical or a
statistical dependence [3, 4]. Some well known
techniques to build this relationship are regression
models [5], statistical approaches [4, 6], radiative
approaches [7], physical-statistical [8], numerical
cloud models [9], artificial neural networks [10]. In
this work, this relationship is modeled through an
artificial neural network approach. In addition, the
improvement of capabilities to develop rain rate
retrieval algorithms, using both artificial neural
networks and data from the Advanced Technology
Microwave Sounder (ATMS), is discussed in this
work.
Some researchers have reported feasibility and
accuracy to estimate rainfall rate from passive
microwave satellites data by using artificial neural
networks [10, 11, 12, 13, 14, 15, 16, 17, 18, 19,
20, 21, 22]. Nevertheless, with the launch of the
Advanced Microwave Sounding Units (AMSU)
sensors, the spectral information in microwave
spectrum about the atmosphere was greatly increased. For instance, the inclusion of channels
centered at 183 GHz in AMSU has provided
invaluable information on the water vapor content,
hydro-meteors concentration, drop liquid water,
vertical atmospheric profile parameters, etc.
Some lessons learned since using AMSU data
to train ANNs to retrieve rain rate are: i) The
development of rain rate retrieval algorithms based
on artificial neural network are suitable when
are combined channels centered at 54 and 183
GHz [18], ii) the 183±7GHz channel shows high
sensitivity to rainfall due to its low sensitivity to
∗
Autor para correspondencia
Correo-e: [email protected] (Erith Muñoz )
surface emission, however in cold and dry regions,
where increase the signal distortion for the surface
emission contribution, the use of a more opaque
channel such as 183±5 GHz or 183±1 GHz trends
to improve the rainfall estimates accuracy [19],
iii) the atmospheric energy budget is affected by
surface emissions, for this reason, to reach an
optimum accuracy level, it is convenient to develop
algorithms for rain rate retrieval over land and
over ocean separately [20], iv) rain rate estimates
over land is less precise than over ocean in the
case of light rainfalls. In contrast, in heavy rainfall
events, the rain rate is underestimate because of
the extinction of radiation caused by the ice cloud
particles of the convective clouds. This interaction
decreases the spaceborne radiometer sensitivity for
detecting cloud liquid water content. [21].
Radiometric and channels features of ATMS are
mostly similar to AMSU, however, the inclusion
of a set of variations represents an important
increase in capabilities because of the possibility
to make observations for 3 additional frequencies
values [23]. In order to retrieve rain rate in heavy
precipitation events, using the Advanced Microwave Technology Sensor (ATMS), a methodology
to design an algorithm able to estimate rainfall
intensities using brightness temperature as data
input is presented in this work. While In this
context, this work deals with two issues, first,
with the analysis of properties of ATMS, in order
to retrieve rain rate by training a set of artificial
neural networks, and second the presentation of a
methodology to design an artificial neural network
to retrieve rain rate using ATMS data.
2. The ANN183 Algorithm
2.1. The Concept
In ANN183, ANN stands for artificial neural
network and 183 for the ATMS channels centered
in the strong absorption line of water vapor at 183
GHz. The development of the ANN183 algorithm
involved two stages, the design and training of the
ANN, and the design of the retrieval algorithm.
The data used to train the neural network was
conformed by simulated brightness temperature
of the ATMS channels and the respectives rain
Revista Ingenierı́a UC, ISSN: 1316–6832, Facultad de Ingenierı́a, Universidad de Carabobo.
E. Muñoz et al / Revista Ingenierı́a UC , Vol. 23, No. 2, agosto 2016, 153-161
155
Figure 1: Flowchart of the ANN183 Algorithm
rate values for that spectral information, for both
summer and winter periods. The data simulation
was carried out by using the radiative transfer
model for the TIROS Operational Vertical Sounder
(RTTOV v-10) [24].
The algorithm is composed by 5 modules,
which can be observed in the Figure 1. The first
module is the ATMS data down-loader procedure,
which is accomplished through the on-line NOAACLASS data manager . NOAA-CLASS facilitates
the choice of data to be downloaded, by taking as
parameters the date and the geographical location
for the acquisition, also it is possible to visualize
the available ATMS scenes before download it.
Once the ATMS data have been downloaded, the
second modules is used to read and setting up
the data to be used as brightness temperature
input into the algorithm. At this point, it is
important to highlight that ATMS data is provided
in HDF5 file format, and the reading and extraction
of data from HDF5 format was performed by
using the open source HDF5 library for C++.
In this step, latitude (lat) and longitude (lon) are
read for each pixel in the scene, and also the
antenna temperature which is expressed as a digital
quantity that must be converted to brightness
temperature magnitude. The third module does the
computation of brightness temperature from the
ATMS data using the equation (1).
!
330
∗ DN,
(1)
T b = 16
2 −9
where T b is the brightness temperature, and DN
is the digital number of the antenna temperature
variable for each ATMS channel. The HDF5 data
format is a 3D cube in which the spatial plane
have dimension of 180x96 pixel, and the z-axis
has 22 layers, one for each ATMS channel. In
this context, when applying equation (1) to ATMS
data, it is possible to built a file conformed by 22
columns, one for each ATMS frequency channel,
and each row in this data represents the brightness
temperature for each pixel in the scene, and for
each frequency. The fourth module consists in the
execution of the artificial neural network. Once the
fourth module is executed, a correction of the rain
rate (RR) output is done. Finally, the fifth and last
module is a map generation tool.
2.2. The Selection Criteria of ATMS Channels
The aim behind the training of artificial neural
networks is to find optimal values for the bias and
synaptic weights to minimize the error function
between the network output and the expected
output (objective) [25]. Since the characteristics
of the network output is highly dependent of its
design, in order to reach convergence and accuracy
in the network output, the channels selected as
input into the network, must be able to predict
the output variable (rain rate). In other words,
each channel assigned to every neuron in the input
layer plays an important role in the generalization
capability of the network, so a prior analysis about
Revista Ingenierı́a UC, ISSN: 1316–6832, Facultad de Ingenierı́a, Universidad de Carabobo.
156
E. Muñoz et al / Revista Ingenierı́a UC , Vol. 23, No. 2, agosto 2016, 153-161
the physical relationship of each ATMS channel
with the rain rate process must be carried out.
Table 1: Central frecuency operational value for the 22
ATMS channels, and espatial resolution features [23]
Channel Frecuency (GHz) Nadir Resolution (km)
1
23,80
75
2
31,40
75
3
50,30
33
4
51,76
33
5
52,80
33
6
53,59
33
7
54,40
33
8
54,94
33
9
55,50
33
10
fLO = 57,29
33
11
fLO ±0,217
33
12
fLO ±0,322±0,048
33
13
fLO ±0,322±0,022
33
14
fLO ±0,322±0,010
33
15
fLO ±0,322±0,045
33
16
89
33
17
166,31
15
18
183,31±7
15
19
183,31±4,5
15
20
183,31±3,0
15
21
183,31±1,8
15
22
183,31±1,0
15
values at which operate each ATMS channel.
The weighting function for each ATMS channel
is shown in Figure 2. In order to choose the
optimal set of channels to be used as data input
into the algorithm, it is important to consider the
sensitivity of each channel to distinguish radiative
parameters related to rainfall activities. Weighting
functions shown that channels 9 to 15 are not
enough sensitive to tropospheric composition,
consequently they are underestimated for rainfall
estimation.
Channels 1 and 2 have been neglected too,
because they provide sparse information about
water vapor and low clouds, as a consequence of
the high sensitivity to radiative surface emissions.
Channels 6, 7 and 8 have been also discarded,
even when they are sensitive to tropospheric
brightness temperature, a bias correction process
should be carry out to remove the dependence on
the scanning angle.
Therefore, the set of channels to be integrated
into the ANN183 algorithm is composed by the
channels 3, 4, 5, 16, 17, 18, 19, 20, 21 and 22. The
aim of this work is to design an algorithm able to
estimate rain rate in severe rainfall events over land
surface. To do this, some experiments for different
combinations of these channels were achieved
[27], and the best results were found when all the
listed channels are combined simultaneously (see
Figure 2).
2.3. Design and Training of the ANN
Figure 2: ATMS weighting functions [26]
In Table 1, are shown the central frecuencies
The main module of the ANN183 algorithm is
the artificial neural network, which must to be
trained to generate a rain rate output with both
satisfactory accuracy and reliability criteria. This
task can be performed by selecting an appropriate
artificial neural network training algorithm, also
an optimum training strategy, providing a training
data set that satisfies the dynamic characteristics of
the phenomenon of precipitation as a function of
the radiative properties that define the brightness
temperature in the atmosphere. It is remarkable also the ANN architecture chosen for the algorithm.
In this context, considering the results reported
by Bellerby [14], about the topology features
Revista Ingenierı́a UC, ISSN: 1316–6832, Facultad de Ingenierı́a, Universidad de Carabobo.
E. Muñoz et al / Revista Ingenierı́a UC , Vol. 23, No. 2, agosto 2016, 153-161
required to emulate the non-linearity of the processes involved in the atmospheric phenomenon
to describe precipitation and its mathematical relationship with radiative transfer variables measured
by satellite sensors, the topology to be used in the
ANN training process are neural networks with
two hidden layer, the first one having 10 neurons
and the second one having 5 neurons. The number
of neurons in the input layer will be the same as the
number of input variables into the algorithm, while
the output layer will be represented by the rain rate
estimation. The data provided into the input layer
is scaled in order to guarantee numerical stability.
The activation function used in the first and second
hidden layer is the hyperbolic tangent, while the
linear activation function was implemented in the
output layer.
The relationship between the rain rate and the
brightness temperature was found through a linear
regression built considering the input and output
layer of the network. For the sake of getting
a satisfactory accuracy in the linear fitting, a
training strategy has been implemented, and the
performance of both L-BFGS and Quasi-Newton
(QN) training methods were evaluated. The criteria to choice these algorithms are based on the
assumption of the stability and the convergence
cost related with algorithms of second orders like
QN method, and taking account that L-BFGS is
considered to be an evolutionary enhancement
of the QN algorithm. Since this assessment, it
is expects to find a sensitive approach in order
to figure out about the performance countenance
provided by recent advances in artificial neural
network algorithms, and their applications to
retrieve rain rate using atmospheric remote sensing
principles [10].
2.4. The Mapping Module
An important task inside the ANN183 algorithm, is the rain rate map generation step. In this
module are plotted both, the rain rate output of
ANN183 for each ATMS scene, and the brightness
temperature for each channel incorporated into the
retrieval algorithm.
157
3. Results
The design of ANN183 derived relevant results
in two aspects: the neural network training, and the
rain rate output found by assess a heavy rainfall
event.
3.1. The Neural Network Training
Table 2: Channels selection for the network training for rain
rate retrieval over land pixels (L). These configurations are
used for both, summer (LS) and winter condition (LW)
Id Experiment Code
Set of Channels
1
LS 1/LW 1
3-4-5-16-17-18-19-20-21-22
2
LS 2/LW 2
18-19-20-21-22
3
LS 3/LW 2
5-17-18-19-20-21-22
The first decision to be taken, in order to configure the network topology, is the ATMS channel
combination to be used as input. Following the
criteria for channels selection discussed early, and
results obtained in prior research [27], in Table 2
are shown the sets of channels considered for the
network training.
Experiment 1 is integrated for all the channels
considered to have both a physical relationship
with precipitation radiative features and no need
for the implementation of an exhaustive correction
step prior to be used as input. Channels for
experiment 2 have been selected due to this combination has reported good results for AMSU data
implementations [5], while the experiment 3 was
done in order to exploit specific features of ATMS
[10, 28]. Each experiment was implemented in an
ANN to be trained by both Quasi-Newton and LBFGS algorithm, in order to looking for the best
output. Results are shown in Table 4, where it can
be observed that the best fitting is for the L-BFGS
algorithm in the experiment 1. The implementation
of the L-BFGS algorithm was possible by using
the ALGLIB library, while QN was implemented
through Open NN [29]. The third column in Table
4, denoted with a percent symbol is the percent
difference in squared-R for both algorithms.
Since results reported in Table 4, we have
selected the channels of the experiment 1 for the
ANN design, and it was trained with the L-BFGS
algorithm.
Revista Ingenierı́a UC, ISSN: 1316–6832, Facultad de Ingenierı́a, Universidad de Carabobo.
158
E. Muñoz et al / Revista Ingenierı́a UC , Vol. 23, No. 2, agosto 2016, 153-161
Table 3: Results summary for the correlation coefficients for
networks trained using the L-BFGS and the Quasi-Newton
(QN) algorithms [27]
Experiment R2 for L-BFGS R2 for QN ( %)
1
0.8391
0.6594
21.41
2
0.4510
0.3397
24.67
3
0.5854
0.3846
34.30
Table 4: Results summary for the correlation coefficients for
networks trained using the L-BFGS and the Quasi-Newton
(QN) algorithms [27]
3.2. Evaluation of a Heavy Rainfall Event
Aimed to evaluate from a qualitative point of
view the output generated for the algorithm, a
heavy rainfall event has been analyzed. On 30
April 2012, a severe rainfall took place over
Kansas, Oklahoma, Missouri and Illinois, in United Stated of America, leading to some large
areas affected by floods due to daily cumulated
precipitation greater than 100 mm, according to
the National Weather Service (NWS) report shown
in Figure 3.a. Furthermore, in Figure 3.b it is
shown the hourly precipitation map from 8:00 to
9:00 UTC, where they were identified three main
precipitating nucleus: i) nucleus 1 was located
between Oklahoma and Texas, with a maximum of
12.5 mm, ii) nucleus 2 located in the boundaries of
Kansas (12.5 mm), Oklahoma (24.5 mm), Missouri
(19.05 mm) and Illinois (19.05 mm), iii) nucleus 3
centered in Indiana with a maximum of 12.5 mm.
The output of ANN183 is shown in the Figure 3.c, where it is remarkable the similarity
between the rainfall pattern reported by NWS
and the output mapped by the ANN183. Main
nucleus identified by the NWS have been also
detected and produced with relevant accuracy. In
relation with the rain-rate estimates values, it is
important to mention that the data provided by the
NWS is the hourly precipitation from 8:00 to 9:00
UTC, while the output generated by ANN183 is
an instantaneous rain-rate. This data was acquired
from measurements carried out at discrete time, for
this reason, a quantitative analysis about the rainrate estimation was dismissed in this work.
However, some key properties to be highlighted
about the results are: 1) Accurate location of the
precipitating cells, 2) clear discrimination between
pixels with low and high precipitation intensity, 3)
non-precipitating pixels classifieds as precipitating
shows intensities lower than 3 mm.
4. Discussion and Conclusion
From a qualitative point of view, the implementation of ANN183 has shown feasibility to evaluate
heavy precipitation events. The implementation of
this algorithm is useful in heavy rainfalls because
the noise could be separated from the precipitation
signal. However, there is still a way to enhance
the accuracy of the algorithm in light rainfall since
the application of training data associated with
strati-form clouds. This could have applications
in the field of data assimilation for early warning
models to floods or hazards related with severe
rainfalls events. ANN183 was designed taking as
reference the ATMS channels weighting functions
in order to exploit the information collected by
each channel in reference to the main radiometric
features of precipitating clouds. The weighting
functions showed that the radiometric characteristics of ATMS involve a set of channels that
provide information about the vertical structures of
clouds, which is an important contribution in order
to estimate rain rate in convective systems.
In this context, the best fit was found when
channels of high (near to 183 GHz) and low
(near to 54 GHz) frequency were combined.
Additionally, the choice of a training algorithm
have a severe impact in the performance and
generalization of the network. Particularly, the LBFGS algorithm provided satisfactory accuracy
levels. In comparison with AMSU, ATMS has
more potential to estimate rain rate due to the
inclusion of channels 19 and 21. The experiment
2 includes all the high frequency channels of
AMSU, and they were not good enough to estimate
the rain rate with a satisfactory accuracy level,
while in the experiment 1, the quality of outputs
is highly related with the inclusion of channels 19
and 21 from ATMS. The number of neurons in the
input of the net is equal to the number of channels
considered to be input into de algorithm, however
the number of hidden layers was in all cases two,
Revista Ingenierı́a UC, ISSN: 1316–6832, Facultad de Ingenierı́a, Universidad de Carabobo.
E. Muñoz et al / Revista Ingenierı́a UC , Vol. 23, No. 2, agosto 2016, 153-161
159
Figure 3: a) The Daily Precipitation map (up) shows the amount of precipitation that was accumulated within 24 hours
[30], b) The hourly precipitation map (down-left) between 08:00 and 09:00 UTC, in which they can be observed three main
precipitating nucleus labeled as 1, 2 and 3 respectively, and c) The instantaneous rain-rate map (down-right) generated by
ANN183 using data captured from 8:00 to 8:08 UTC. Note that difference between the NWS hourly cumulated precipitation
and the rain rate output of ANN183 are expected, and for this reason, a quantitative comparison between them would be
ambiguous. The rain rate ANN183 output shows a noisy signal with an intensity less 5mm. Even when this noise could be
filtered, this bias limits the accuracy of this algorithm to classify as precipitating pixels in a strati-form clouds cover.
with 10 neurons the first and 5 the second.
The instantaneous rain rate estimates done by
ANN183 showed good performance to generate
the precipitation intensities pattern reported by
NWS. Also, the deeper convective cloud cells
were associated with the highest rain rate in the
scene. It is to be expected an underestimation of
rainfall intensities related with convective clouds,
however the implementation of a calibration step
could decrease the difference in estimates. Further
research are been developed in order to calibrate
the output for intensities, and validation of the
methodology.
Finally, the training data involved radiative
conditios for summer and winter periods, so this
algorithm have been developed for mid latitudes.
In the case of lower latitudes, the training of
an ANN using data more associated with the
radiative properties of these environments could be
necessary.
Agradecimientos
El autor agradece al Instituto de Altos Estudios
Espaciales Mario Gulich, de la Comisión Nacional
de Actividades Espaciales de Argentina por el
soporte y financiamiento para llevar a cabo esta investigación. De igual forma, agradece la gentileza
del Dr. Vincenzo Levizzani del CNR de Italia, y a
su grupo de colaboradores, por el entrenamiento
y transferencia de conocimientos en los tópicos
discutidos en este trabajo.
Referencias
[1] Henry E Fleming, Norman C Grody, and Eugene J
Kratz. The forward problem and corrections for
the ssm/t satellite microwave temperature sounder.
Geoscience and Remote Sensing, IEEE Transactions
on, 29(4):571–583, 1991.
[2] Christian Kummerow, William S Olson, and Louis
Giglio. A simplified scheme for obtaining precipitation
and vertical hydrometeor profiles from passive micro-
Revista Ingenierı́a UC, ISSN: 1316–6832, Facultad de Ingenierı́a, Universidad de Carabobo.
160
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
E. Muñoz et al / Revista Ingenierı́a UC , Vol. 23, No. 2, agosto 2016, 153-161
wave sensors. Geoscience and Remote Sensing, IEEE
Transactions on, 34(5):1213–1232, 1996.
Nazzareno Pierdicca, Frank Silvio Marzano, Giovanni D Auria, Patrizia Basili, Piero Ciotti, and Alberto
Mugnai.
Precipitation retrieval from spaceborne
microwave radiometers based on maximum aa posteriori probability estimation. Geoscience and Remote
Sensing, IEEE Transactions on, 34(4):831–846, 1996.
G d’Auria, FS Marzano, N Pierdicca, P Basili, and
P Ciotti.
Rain retrieval algorithms for passive
microwave observations: a comparison and a choice.
In Geoscience and Remote Sensing Symposium, 1995.
IGARSS’95.’Quantitative Remote Sensing for Science
and Applications’, International, volume 2, pages
1136–1138. IEEE, 1995.
S. Laviola and V Levizzani.
The 183-wsl fast
rain rate retrieval algorithm. part 1: Retrieval design.
Atmospheric Research, pages 443–461, 2011.
S Di Michele, A Tassa, A Mugnai, and FS Marzano.
The bayesian algorithm for microwave precipitation
retrieval (bampr): potential and application to trmm
data. In Geoscience and Remote Sensing Symposium,
2003. IGARSS’03. Proceedings. 2003 IEEE International, volume 5, pages 3142–3144. IEEE, 2003.
Frank Silvio Marzano, Ermanno Fionda, Piero Ciotti,
and Antonio Martellucci. Ground-based multifrequency microwave radiometry for rainfall remote
sensing.
Geoscience and Remote Sensing, IEEE
Transactions on, 40(4):742–759, 2002.
Luca Pulvirenti, Nazzareno Pierdicca, Frank Silvio
Marzano, Paolo Castracane, and Giovanni D Auria. A
physical-statistical approach to match passive microwave retrieval of rainfall to mediterranean climatology.
Geoscience and Remote Sensing, IEEE Transactions
on, 40(10):2271–2284, 2002.
Chinnawat Surussavadee and David H Staelin. Global satellite millimeter-wave precipitation retrievals
trained with a cloud-resolving numerical weather
prediction model. In Geoscience and Remote Sensing
Symposium, 2007. IGARSS 2007. IEEE International,
pages 3910–3913. IEEE, 2007.
Erith Alexander Munoz, Francesco Di Paola, and
Mario Lanfri. Advances on rain rate retrieval from
satellite platforms using artificial neural networks.
Latin America Transactions, IEEE (Revista IEEE
America Latina), 13(10):3179–3186, 2015.
S. Lovejot and G.L. Austin. The delineation of rain
areas from visible and infrared satellite data for gate
and midlatitudes. Atmos. Ocean., pages 1048–1054,
1979.
K. Hsu, X. Gao, S. Sorooshian, and .V. Gupta. Precipitation estimation from remotely sensed information
using artificial neural networks. J. Appl. Meteor., pages
1176–1190, 1997.
D. Tsintikidis, J.L. Haferman, and N. Anagnostou. A
neural network approach to estimating rainfall from
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
spaceborne microwave data. IEEE Trans. Geosci.
Remote Sens., pages 1079–1092, 1997.
T. Bellerby, M. Todd, D. Kniveton, and C. Kidd.
Rainfall estimation from a combination of trmm
precipitation radar and goes multispectral satellite
imagery through the use of an artificial neural network.
American Meteorological Society, pages 2115–2128,
2000.
D.I.F. Grimes, E. Coppola, M. Verdecchia, and
G. Visconti. A neural network approach to realtime rainfall estimation for africa usind satellite data.
american Meteorological Society, pages 1119–1133,
2003.
F. J. Tapiador, C. Kidd, K.L. Hsu, and F. Marzano.
Neural networks in satellite rainfall estimation. Meteorol. Appl., pages 83–91, 2004.
F.W. Chen and D.H. Staelin. Airs/amsu/hsb precipitation estimates. IEEE Transactions on Geosciencie and
Remote Sensing, pages 410–417, 2003.
D. H. Staelin and F.W. Chen. precipitation observations
near 54 and 183 ghz using the nooa-15 satellite. IEEE
Transactions on Geosciencie and Remote Sensing,
pages 2322–2332, 2000.
W. J. Blackwell and F.W. Chen. Neural network
applications in high-resolution atmospheric remote
sensing. Lincoln Laboratory Journal, pages 299–322,
2005.
Chinnawat Surussavadee and David H Staelin. Global
millimeter-wave precipitation retrievals trained with a
cloud-resolving numerical weather prediction model,
part i: Retrieval design. Geoscience and Remote
Sensing, IEEE Transactions on, 46(1):99–108, 2008.
Chinnawat Surussavadee and David H Staelin. Global
millimeter-wave precipitation retrievals trained with a
cloud-resolving numerical weather-prediction model,
part ii: Performance evaluation. Geoscience and
Remote Sensing, IEEE Transactions on, 46(1):109–
118, 2008.
R. V. Leslie, W. J. Blackwell, L. J. Bickmeier, and
L. G. Jairam. Neural network microwave precipitation
retrievals and modeling results. In Microwave Remote
Sensing of the Atmosphere and Environment VI, 2008.
Ch. Surussavadee and D.H. Staelin. Npoess precipitation retrievals using the atms passive microwave
spectrometer. IEEE Geosciencie and Remote Sensing
Letters, pages 440–444, 2010.
R. Saunders, J. Hocking, P. Rayer, M. Matricardi,
A. Geer, N. Bormann, P. Brunel, F. Karbou, and
F. Aires. Rttov-10 science and validation report.
EUMETSAT, 2006.
Leandro Nunes De Castro. Fundamentals of natural
computing: basic concepts, algorithms, and applications. CRC Press, 2006.
John A Knaff, Raymond M Zehr, Mitchell D Goldberg,
and Stanley Q Kidder. An example of temperature
structure differences in two cyclone systems derived
Revista Ingenierı́a UC, ISSN: 1316–6832, Facultad de Ingenierı́a, Universidad de Carabobo.
E. Muñoz et al / Revista Ingenierı́a UC , Vol. 23, No. 2, agosto 2016, 153-161
[27]
[28]
[29]
[30]
from the advanced microwave sounder unit. Weather
and forecasting, 15(4):476–483, 2000.
Erith Muñoz and Cesar Seijas. Análisis comparativo de
ajuste en entrenamiento de redes neuronales artificiales
a partir de las librerı́as open nn y alglib. La Granja,
21(1), 2015.
Erith Alexander Munoz, Francesco Di Paola, Mario
Lanfri, and Francisco J Arteaga. Observing the troposphere through the advanced technology microwave
sensor (atms) to retrieve rain rate. Latin America
Transactions, IEEE (Revista IEEE America Latina),
14(2):586–594, 2016.
R. López. Open nn: An open source neural networks
c++ library. Technical report, Open Neural Network
Library, 2014. http://www.cimne.com/flood/.
John McEnery, John Ingram, Qingyun Duan, Thomas
Adams, and Lee Anderson.
Noaa’s advanced
hydrologic prediction service. Bulletin of the American
Meteorological Society, 86(3):375, 2005.
Revista Ingenierı́a UC, ISSN: 1316–6832, Facultad de Ingenierı́a, Universidad de Carabobo.
161